Advancements in QSAR modelling: Decision trees and rotation forest for prediction of Aspergillus anti-inflammatory metabolites

This study presents applications of advancements in QSAR modelling for predicting nitric oxide (NO) inhibitors and anti-inflammatory metabolites from the <em>Aspergillus</em> genus. Inflammation-related diseases remain a pressing concern, necessitating the identification of effective anti-inflammatory compounds. Using decision trees, the Ranker method, and CorrelationAtrributeEval as a base classifier for attribute selection together with Rotation Forest and Adaboost as enhancers, we explored their potential with different classifiers including Artificial Neural Networks and J48 Trees. The proposed QSAR models employed an ensemble approach with Rotation Forest and Adaboost.M1, applying an automated KNIME workflow. Seven molecular descriptors were selected and trained on a comprehensive dataset of diverse anti-inflammatory <em>Aspergillus</em> specialised metabolites. Results showed that the Rotation Forest-enhanced version outperformed other models, capturing complex structure-activity relationships and improving predictive performance. Chemical characteristics of electrotopological state, topological distances, and functional groups including secondary amides and alcohols contribute to important anti-inflammatory effects. The developed QSAR model showed good predictive performance for anti-inflammatory <em>Aspergillus</em> metabolites, focusing on their NO inhibitory activity. These results can contribute to the discovery of novel anti-inflammatory drugs based on computational techniques.本研究介绍了定量构效关系（QSAR, Quantitative Structure-Activity Relationship）建模进展的应用，用于预测曲霉属（Aspergillus）来源的一氧化氮（NO, Nitric Oxide）抑制剂与抗炎代谢物。炎症相关疾病仍是亟待解决的突出公共健康问题，亟需筛选高效抗炎活性化合物。本研究以决策树（Decision Trees）、Ranker法（Ranker method）与相关属性评估法（CorrelationAttributeEval）作为属性选择的基础分类器，并结合旋转森林（Rotation Forest）与自适应提升（Adaboost）作为增强算法，同时探索了人工神经网络（Artificial Neural Networks）、J48树（J48 Trees）等不同分类器的应用潜力。所提出的QSAR模型采用融合旋转森林与Adaboost.M1的集成学习框架，并依托自动化KNIME工作流（KNIME workflow）完成建模流程。研究选取7种分子描述符，在涵盖多样抗炎曲霉属专属代谢物的综合数据集上开展模型训练。实验结果显示，经旋转森林增强的模型性能优于其余所有模型，能够有效捕捉复杂的构效关系并提升预测表现。电拓扑状态、拓扑距离以及二级酰胺、醇类等官能团的化学特征，对抗炎活性具有关键贡献。所开发的QSAR模型针对具备一氧化氮抑制活性的曲霉属抗炎代谢物展现出优异的预测性能。本研究成果可为基于计算技术的新型抗炎药物研发提供有力支撑。